A rigorous verification protocol that mandates fresh, objective evidence before an AI agent can claim task completion or perform a commit.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install verify-before-done
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install verify-before-done using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Verify Before Done?
The Verify Before Done skill acts as a critical gatekeeper for AI-driven development. It is designed to prevent false completion claims by requiring agents to provide verifiable proof—such as test logs, linter outputs, or exit codes—before finalizing any work. By integrating this into your Openclaw Skills library, you ensure that every status report, push, or PR is backed by actual execution rather than assumptions or previous state data.
This skill addresses the common issue of hallucinated success in AI agents. It forces a mindset shift where the agent must identify the specific proof required, execute the necessary commands, and validate the output against strict success criteria. In the ecosystem of Openclaw Skills, this protocol is essential for maintaining high-quality codebases and building trust between human developers and AI agents.
Verify Before Done Use Cases
- Before making any git commit or pushing code to a remote repository.
- Validating that a bug fix actually resolves the original issue using reproduction scripts.
- Ensuring 0 linter errors and 0 test failures before submitting a Pull Request.
- Verifying the output of sub-agents or delegated tasks to ensure they meet project standards.
- Confirming that all items in a requirement checklist have been physically implemented and tested.
How Verify Before Done Works
- Identify: Determine which specific command or check will provide undeniable proof of the current status.
- Execute: Run the identified command immediately to capture fresh output; using historical logs is strictly prohibited.
- Read: Analyze the complete output, including standard error (stderr), exit codes, and failure counts.
- Verify: Compare the results against the original claim. If the output confirms the claim, proceed; if not, report the actual state with evidence.
- Claim: Only after successful verification can the agent issue a completion claim or move to the next workflow step.
Verify Before Done Setup
To integrate this protocol into your Openclaw Skills workflow, include the skill file in your agent's accessible skills directory:
# Navigate to your agent's skill directory
cd path/to/your/agent/skills
# Clone or copy the verify-before-done.md file into the directory
cp ~/downloads/verify-before-done.md .
Ensure your system prompt instructs the agent to consult this skill before any operation involving the terms 'done', 'fixed', 'implemented', or 'complete'.
Verify Before Done Data Schema & Taxonomy
The skill utilizes a structured verification matrix to determine the validity of a claim based on evidence types:
| Claim Type | Required Evidence | Insufficient Evidence |
|---|---|---|
| Test Passage | Command output showing 0 failures | "It should pass" or old logs |
| Linter Status | Linter output showing 0 errors | Partial checks or assumptions |
| Build Success | Build command exiting with code 0 | "Logs look fine" without exit check |
| Bug Fix | Passing reproduction test | Code changes without verification |
| Requirements | Itemized checklist with proof | General success claims |
Verify Before Done Advanced Features
- Stop-Word Triggering: Automatically pauses execution when speculative language like "probably", "I think", or "should work" is detected.
- Sub-agent Audit: Provides a framework for main agents to independently verify the work reported by sub-agents using Openclaw Skills.
- VCS Integration: Mandates a git diff check to confirm that intended changes actually exist on the filesystem before verification.
- Zero-Tolerance Exception Logic: Removes the ability for agents to bypass verification due to fatigue, complexity, or previous successful runs.
SKILL.md
Loading
Related Openclaw Skills
TDD Discipline
A rigorous Test-Driven Development framework for AI agents to ensure code reliability and eliminate regressions.
Subagent-Driven Development (subagent-dev)
A structured implementation pattern that uses fresh subagents for every task to ensure spec compliance and high code quality.
Stripe Payments Integration
A professional integration guide for implementing secure, modern Stripe payment flows including subscriptions and marketplaces.
Sprite Sheet & Texture Atlas Mastery
A technical guide and toolkit for optimizing 2D game assets using sprite sheets and texture atlases within Rust and Godot environments.
Video Pro
A high-performance video editing skill leveraging Remotion and FFmpeg for programmatic video production and processing on remote nodes.
Web Bundling (Single HTML)
A specialized workflow for bundling modern React and Vite applications into portable, single-file HTML artifacts for seamless distribution.
Featured*
